Anthropic's browseragent werd in 31,5% van de gevallen gehackt voordat beveiligingen ingrepen

Anthropic disclosed that attackers successfully hijacked its browser agent in 31.5% of test cases before safety measures activated, making it the only major AI lab to publish specific prompt injection vulnerability data this spring. The transparency contrasts with OpenAI, Google, and Meta, which have not released comparable security metrics for their systems.